Abstract
Observational studies differ from experiments in that randomization is not used to assign treatments. How were treatments assigned? This chapter introduces two simple models for treatment assignment in observational studies. The first model is useful but naïve: it says that people who look comparable are comparable. The second model speaks to a central concern in observational studies: people who look comparable in the observed data may not actually be comparable; they may differ in ways we did not observe.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Similar content being viewed by others
Notes
- 1.
The alert reader will notice that the supposition that there could be two subjects, k and ℓ, with Z k + Z ℓ = 1 but π k = π ℓ is already a step beyond the mere representation in (3.1), because it precludes setting u i = Z i for all i and instead requires 0 < π i < 1 for at least two subjects.
- 2.
- 3.
The argument just presented is the simplest version of a general argument developed in [81].
- 4.
See, for instance, the argument made by David Gross and Nicholas Souleles, quoted in Sect. 4.2 from their paper [40]. The example in Chap. 8 attempts to use both approaches, although it is, perhaps, not entirely convincing in either attempt. In Chap. 8 , the dose of chemotherapy received by patients with ovarian cancer is varied by comparing two types of providers of chemotherapy, gynecological oncologists (GO) and medical oncologists (MO), where the latter treat more intensively. The most important clinical covariates are measured, including clinical stage, tumor grade, histology, and comorbid conditions. Despite these two good attempts, it is far from clear that the GO-vs-MO assignment is haphazard, even after adjustments, and it is quite likely that the oncologist knew considerably more about the patient than is recorded in electronic form.
- 5.
Figure 1.2 is an example of quasi-experimental reasoning. Instead of comparing college attendance in two groups, say the children of deceased fathers in 1979–1981, when the benefit program was available, and the children of deceased fathers in 1982–1983, after the program was withdrawn, Fig. 1.2 follows Susan Dynarski’s [26] study in comparing four groups. Figure 1.2 shows at most a small shift in college attendance by the group of children without deceased fathers, who were never eligible for the program. This is a small check, but a useful one: it suggests that the comparison of children of deceased fathers in 1979–1981 and 1982–1983, though totally confounded with time, may not be seriously biased by this confounding, because a substantial shift in college attendance by unaffected groups over this time is not evident. The quasi-experimental design in Fig. 1.2 is widely used and is known by a variety of not entirely helpful names, including “pretest–posttest nonequivalent control group design” and “difference-in-differences design.” Kevin Volpp et al. [131] used this design in a medical context to study the possible effects on patient mortality of a reduction, to 80 h per week, of the maximum number of hours that a medical resident may work in a hospital. See [1, 7, 9, 14, 92] for discussion in general terms. The study by Garen Wintemute et al. [135], as discussed in Sect. 4.3, contains another example of quasi-experimental reasoning using a “control construct,” that is, an outcome thought to be unaffected by the treatment. They were interested in the possible effects of a change in gun control laws on violent crimes using guns. One would not expect a change in gun control laws to have a substantial effect on nonviolent, nongun crimes, and so the rate of such crimes is their control construct or unaffected outcome. See Sect. 4.3. An example with two control groups subject to different biases is discussed in Sect. 12.3.
- 6.
Technically, strongly ignorable treatment assignment given x is \( \left . Z\, \underline {\;\left \vert \;\right \vert \;}\,\left ( r_{T},\,r_{C}\right ) \,\right \vert \,\mathbf {x}\) and \(0<\Pr \left ( \left . Z=1\,\right \vert \,\, \mathbf {x}\right ) <1\) for all x, or in a single expression, \( 0<\Pr \left ( \left . Z=1\,\right \vert \,r_{T},\,r_{C},\,\mathbf {x}\right ) \) \( =\Pr \left ( \left . Z=1\,\right \vert \,\,\mathbf {x}\right ) <1\) for all x; that is, no explicit reference is made to an unobserved covariate u. I introduce u here, rather than in Sect. 3.4 where it is essential, to simplify the transition between Sects. 3.3 and 3.4.
The remainder of this footnote is a slightly technical discussion of the formal relationship between (3.5) and the condition as given in [104]; it will interest only the insanely curious. In a straightforward way [23, Lemma 3] , condition (3.5) implies \(\left . Z\, \underline { \;\left \vert \;\right \vert \;}\,\left ( r_{T},\,r_{C}\right ) \,\right \vert \, \mathbf {x}\) and \(0<\Pr \left ( \left . Z=1\,\right \vert \,\,\mathbf {x}\right ) <1\) for all x. Also, in a straightforward way [23, Lemma 3] , if (3.5) is true then (i) \(\left . Z\, \underline { \;\left \vert \;\right \vert \;}\,\left ( r_{T},\,r_{C}\right ) \,\right \vert \,\left ( \mathbf {x},u\right ) \), (ii) \(0<\Pr \left ( \left . Z=1\,\right \vert \,\,\mathbf {x}\right ) <1\) for all x, and (iii) \(\left . Z\, \underline {\;\left \vert \;\right \vert \;}u\,\right \vert \,\mathbf {x}\) are true. Moreover, (ii) and (iii) together imply (iv) \(0<\Pr \left ( \left . Z=1\,\right \vert \,\,\mathbf {x},u\right ) <1\) for all \(\left ( \mathbf {x} ,u\right ) \). In words, condition (3.5) implies both strong ignorability given x and also strong ignorability given \(\left ( \mathbf {x},u\right ) \). Now, conditions (i) and (iv), without condition (iii), are the key elements of the sensitivity model in Sect. 3.4: they say that treatment assignment would have been strongly ignorable if u had been measured and included with x; so the failure to measure u is the source of our problems. Moreover, if condition (iii) were true in addition to (i) and (iv), then (3.5) and strong ignorability given x follow. In brief, strong ignorability given x is implied by the addition of condition (iii) to strong ignorability given \(\left ( \mathbf {x} ,u\right ) \). In words, one can reasonably think of the naive model Sect. 3.3 as the sensitivity model of Sect. 3.4 together with the irrelevance of u as expressed by \(\left . Z\, \underline {\;\left \vert \;\right \vert \;}u\,\right \vert \, \mathbf {x}\) in condition (iii).
- 7.
The proof is easy [104]. Recall the following basic facts about conditional expectations: if A, B, and C are random variables, then \(E\left ( A\right ) =E\left \{ E\left ( \left . A \mathbf {\,}\right \vert \,\,B\,\right ) \right \} \) and \(E\left ( \left . A \mathbf {\,}\right \vert \,\,C\,\right ) =E\left \{ \left . E\left ( \left . A \mathbf {\,}\right \vert \,\,B,C\,\right ) \,\right \vert \,C\right \} \). From the definition of conditional independence, to prove (3.10), it suffices to prove \(\Pr \left \{ \left . Z=1 \mathbf {\,}\right \vert \,\,\mathbf {x},\,e\left ( \mathbf {x}\right ) \right \} =\Pr \left \{ \left . Z=1\mathbf {\,}\right \vert \,\,\,e\left ( \mathbf {x} \right ) \right \} \). Because \(e\left ( \mathbf {x}\right ) \) is a function of x, conditioning on x fixes \(e\left ( \mathbf {x}\right ) \) , so \(\Pr \left \{ \left . Z=1\mathbf {\,}\right \vert \,\,\mathbf {x},\,e\left ( \mathbf {x}\right ) \right \} \) \(=\Pr \left ( \left . Z=1\mathbf {\,}\right \vert \,\,\mathbf {x}\right ) =e\left ( \mathbf {x}\right ) \). Because Z = 0 or Z = 1 , for any random variable D, we have \(\Pr \left ( \left . Z=1\mathbf {\,} \right \vert \,\,D\right ) =E\left ( \left . Z\mathbf {\,}\right \vert \,D\right ) \) . With A = Z, B = x, \(C=e\left ( \mathbf {x}\right ) \), we have \(\Pr \left \{ \left . Z=1\mathbf {\,}\right \vert \,\,\,e\left ( \mathbf {x}\right ) \right \} =E\left \{ \left . Z\mathbf {\,}\right \vert \,\,e\left ( \mathbf {x} \right ) \,\right \} \) \(=E\left \{ \left . E\left \{ \left . Z\mathbf {\,} \right \vert \,\,\mathbf {x},\,e\left ( \mathbf {x}\right ) \right \} \,\right \vert \,e\left ( \mathbf {x}\right ) \right \} \) \(=E\left \{ \left . \Pr \left \{ \left . Z=1\mathbf {\,}\right \vert \,\,\mathbf {x},\,e\left ( \mathbf {x} \right ) \right \} \,\right \vert \,e\left ( \mathbf {x}\right ) \right \} \) \( =E\left \{ \left . e\left ( \mathbf {x}\right ) \,\,\right \vert \,e\left ( \mathbf { x}\right ) \right \} \) \(=e\left ( \mathbf {x}\right ) =\Pr \left ( \left . Z=1 \mathbf {\,}\right \vert \,\,\mathbf {x}\right ) \), which is what we needed to prove.
- 8.
Part II makes use of a stronger version of the balancing property than stated in (3.10). Specifically, if you match on \( e\left ( \mathbf {x}\right ) \) and any other aspect of x, say \( h\left ( \mathbf {x}\right ) \), you still balance x, that is, \( \left . Z\, \underline {\;\left \vert \;\right \vert \;}\,\mathbf {x}\,\right \vert \,\left \{ e\left ( \mathbf {x}\right ) ,\,h\left ( \mathbf {x}\right ) \right \} \) for any \(h\left ( \mathbf {x}\right ) \). See [104] for the small adjustments required in the proof.
- 9.
- 10.
In technical terms, if treatment assignment is strongly ignorable given x, then it is strongly ignorable given \(e\left ( \mathbf {x}\right ) \) . The proof of this version is also short [104]. In footnote 7, we saw that \(\Pr \left \{ \left . Z=1\mathbf {\, }\right \vert \,\,\,e\left ( \mathbf {x}\right ) \right \} =\Pr \left ( \left . Z=1 \mathbf {\,}\right \vert \,\,\mathbf {x}\right ) \). Condition (3.11) says \(\Pr \left ( \left . Z=1\,\right \vert \,r_{T},\,r_{C},\,\mathbf {x},\,u\right ) =\Pr \left ( \left . Z=1\,\right \vert \,\,\mathbf {x}\right ) \). Together they say \(\Pr \left ( \left . Z=1\,\right \vert \,r_{T},\,r_{C},\,\mathbf {x},\,u\right ) =\Pr \left \{ \left . Z=1\mathbf {\,}\right \vert \,\,\,e\left ( \mathbf {x}\right ) \right \} \), which implies (3.12).
- 11.
In general, a sensitivity analysis asks how the conclusion of an argument dependent upon assumptions would change if the assumptions were relaxed. The term is sometimes misused to refer to performing several parallel statistical analyses without regard to the assumptions upon which they depend. If several statistical analyses all depend upon the same assumption—for instance, the naïve model (3.5)—then performing several such analyses provides no insight into consequences of the failure of that assumption.
- 12.
Odds are an alternative way of expressing probabilities. Probabilities and odds carry the same information in different forms. A probability of π k = 2∕3 is an odds of \(\pi _{k}/\left ( 1-\pi _{k}\right ) =2\) or 2-to-1. Gamblers prefer odds to probabilities because odds express the chance of an event in terms of fair betting odds, the price of a fair bet. It is easy to move from probability π k to odds \(\omega _{k}=\pi _{k}/\left ( 1-\pi _{k}\right ) \) and back again from odds ω k to probability \(\pi _{k}=\omega _{k}/\left ( 1+\omega _{k}\right ) \).
- 13.
Implicitly, the critic is saying that the failure to measure u is the source of the problem, or that (3.5) would be true with \(\left ( \mathbf {x} ,u\right ) \) in place of x, but is untrue with x alone. That is, the critic is saying \(\pi _{\ell }=\Pr \left ( \left . Z_{\ell }=1\,\right \vert \,r_{T\ell },\,r_{C\ell },\,{\mathbf {x}}_{\ell },\,u_{\ell }\right ) \) \(=\Pr \left ( \left . Z_{\ell }=1\,\right \vert \,{\mathbf {x}}_{\ell },\,u_{\ell }\right ) \). As in Sect. 3.1, because of the delicate nature of unobserved variables, this is a manner of speaking rather than a tangible distinction. If the formalities are understood to refer to \(\pi _{\ell }=\Pr \left ( \left . Z_{\ell }=1\,\right \vert \,r_{T\ell },\,r_{C\ell },\,{\mathbf {x}}_{\ell },\,u_{\ell }\right ) \), then it is not necessary to insist that \(\pi _{\ell }=\Pr \left ( \left . Z_{\ell }=1\,\right \vert \,\,{\mathbf {x}}_{\ell },\,u_{\ell }\right ) \). Conversely, there is always a scalar unobserved covariate u with 0 ≤ u ≤ 1 such that \(\left . Z\, \underline { \;\left \vert \;\right \vert \;}\,\left ( r_{T},\,r_{C}\right ) \,\right \vert \,\left ( \mathbf {x},u\right ) \), namely the unobserved covariate u = ζ, where \(\zeta =\Pr \left ( Z=1 | r_{T}, r_{C}, \mathbf {x} \right )\); for proof, see [102, §6.3]. Consistent with terminology introduced by Frangakis and Rubin [33], we might call this unobserved covariate, u = ζ, the “principal unobserved covariate.” In effect, the bounded, scalar principal unobserved covariate, u = ζ, is the only unobserved covariate that matters, and it always exists. A finite Γ changes the principal unobserved covariate, u = ζ, from a representation to a model, because a finite Γ implies 0 < ζ < 1 rather than 0 ≤ ζ ≤ 1.
- 14.
In a fussy technical sense, the numbering of pairs and people within pairs is supposed to convey nothing about these people, except that they were eligible to be paired, that is, they have the same observed covariates, different treatments, with 2I distinct people. Information about people is supposed to be recorded in variables that describe them, such as Z, x, u, r T, r C, not in their position in the data set. You cannot put your brother-in-law in the last pair just because of that remark he made last Thanksgiving; you have to code him in an explicit brother-in-law variable. Obviously, it is easy to make up subscripts that meet this fussy requirement: number the pairs at random, then number the people in a pair at random. The fussy technical point is that, in going from the L people in (3.1) to the 2I paired people, no information has been added and tucked away into the subject numbers—the criteria for pairs are precisely x i1 = x i2 , Z i1 + Z i2 = 1 with 2I distinct individuals.
- 15.
What does it mean to speak of the “largest distribution”? One random variable, A, is said to be stochastically larger than another random variable, B, if \(\Pr \left ( A\geq k\right ) \geq \Pr \left ( B\geq k\right ) \) for every k. That is, A is more likely than B to jump over a bar at height k, no matter how high, k, the bar is set. Because this must be true for every k, it is a rather special relationship between random variables. For instance, it might happen that \(\Pr \left ( A\geq -1.65\right ) =0.95\), \(\Pr \left ( A\geq 1.65\right ) =0.05\) while \(\Pr \left ( B\geq -1.65\right ) =0.80\), \(\ \Pr \left ( B\geq 1.65\right ) =0.20\), so neither A nor B is stochastically larger than the other. For instance, this is true if A is Normal with mean zero and standard deviation 1, and B is Normal with mean zero and standard deviation 2. The intuition is overwhelmingly strong that the largest null distribution of Wilcoxon’s signed rank statistic is obtained by driving the chance of a positive difference up to its maximum, namely \(\varGamma /\left ( 1+\varGamma \right ) \) in (3.18), and this intuition turns out to be correct. It takes a small amount of effort in this case to show the distribution is actually stochastically largest; see [93, Chapter 4] . In other cases, the same result is available with somewhat more effort. In still other cases, one can speak of a “largest distribution” only asymptotically, that is, only in large samples; see [37, 101] or [93, Chapter 4]. This presents no problem in practice, because the asymptotic results are quite adequate and easy to use [93, Chapter 4] ; however, it does make the theory of the paired case simpler than, say, the theory of matching with several controls matched to each treated subject. To see a parallel discussion of the paired case and the case of several controls, see [97]. For technical details, see [101].
- 16.
- 17.
Bibliography
Abadie, A. : Semiparametric difference-in-differences estimators. Rev. Econ. Stud. 72, 1–19 (2005)
Albers, W., Bickel, P.J., van Zwet, W.R.: Asymptotic expansions for the power of distribution free tests in the one-sample problem. Ann. Stat. 4, 108–156 (1976)
Angrist, J. , Hahn, J. : When to control for covariates? Panel asymptotics for estimates of treatment effects. Rev. Econ. Stat. 86, 58–72 (2004)
Angrist, J.D., Krueger, A.B.: Empirical strategies in labor economics. In: Ashenfelter, O., Card, D. (eds.) Handbook of Labor Economics, vol. 3, pp. 1277–1366. Elsevier, New York (1999)
Angrist, J.D., Lavy, V.: Using Maimonides’ rule to estimate the effect of class size on scholastic achievement. Q. J. Econ. 114, 533–575 (1999)
Armstrong, C.S., Blouin, J.L., Larcker, D.F.: The incentives for tax planning. J. Accounting Econ. 53, 391–411 (2012)
Athey, S., Imbens, G.W.: Identification and inference in nonlinear difference-in-differences models. Econometrica 74, 431–497 (2006)
Becker, S.O., Caliendo, M.: Sensitivity analysis for average treatment effects. Stata J. 7, 71–83 (2007)
Bertrand, M., Duflo, E., Mullainathan, S.: How much should we trust difference-in-differences estimates? Q. J. Econ. 119, 249–275 (2004)
Besley, T., Case, A.: Unnatural experiments? Estimating the incidence of endogenous policies. Econ. J. 110, 672–694 (2000)
Bross, I.D.J.: Statistical criticism. Cancer 13, 394–400 (1961)
Bross, I.D.J.: Spurious effects from an extraneous variable. J. Chron. Dis. 19, 637–647 (1966)
Campbell, D.T.: Factors relevant to the validity of experiments in social settings. Psychol. Bull. 54, 297–312 (1957)
Campbell, D.T.: Reforms as experiments. Am. Psychol. 24, 409–429 (1969)
Campbell , D.T.: Methodology and Epistemology for Social Science: Selected Papers. University of Chicago Press, Chicago (1988)
Carnegie, N.B., Harada, M. Hill, J.L.: Assessing sensitivity to unmeasured confounding using a simulated potential confounder. J. Res. Educ. Effect 9, 395–420 (2016)
Chi, S.S., Shanthikumar, D.M.: Local bias in Google search and the market response around earnings announcements. Account Rev. 92, 115–143 (2016)
Copas, J.B., Eguchi, S.: Local sensitivity approximations for selectivity bias. J. R. Stat. Soc. B 63, 871–896 (2001)
Copas, J.B., Li, H.G.: Inference for non-random samples. J. R. Stat. Soc. B 59, 55–77 (1997)
Cornfield, J. , Haenszel, W. , Hammond, E. , Lilienfeld, A. , Shimkin, M. , Wynder, E. : Smoking and lung cancer: recent evidence and a discussion of some questions. J. Natl. Cancer Inst. 22, 173–203 (1959)
Crump, R.K., Hotz, V.J., Imbens, G.W., Mitnik, O.A.: Dealing with limited overlap in estimation of average treatment effects. Biometrika 96, 187–199 (2009)
D’Agostino, R.B.: Propensity score methods for bias reduction in the comparison of a treatment to a non-randomized control group. Stat. Med. 17, 2265–2281 (1998)
Dawid, A.P.: Conditional independence in statistical theory (with Discussion). J. R. Stat. Soc. B 41, 1–31 (1979)
Dehejia, R.H., Wahba, S.: Propensity score-matching methods for nonexperimental causal studies. Rev. Econ. Stat. 84, 151–161 (2002)
Diprete, T.A., Gangl, M.: Assessing bias in the estimation of causal effects: Rosenbaum bounds on matching estimators and instrumental variables estimating with imperfect instruments. Sociol. Method 34, 271–310 (2004)
Dynarski, S.M.: Does aid matter? Measuring the effect of student aid on college attendance and completion. Am. Econ. Rev. 93, 279–288 (2003)
Fenech, M., Changb, W.P., Kirsch-Voldersc, M., Holland, N., Bonassie, S., Zeiger, E.: HUMN project: detailed description of the scoring criteria for the cytokinesis-block micronucleus assay using isolated human lymphocyte cultures. Mutat. Res. 534, 65–75 (2003)
Fogarty, C.B.: Studentized sensitivity analysis for the sample average treatment effect in paired observational studies. J. Am. Stat. Assoc. (2019, to appear). https://doi.org/10.1080/01621459.2019.1632072
Fogarty, C.B., Small, D.S.: Sensitivity analysis for multiple comparisons in matched observational studies through quadratically constrained linear programming. J. Am. Stat. Assoc. 111, 1820–1830 (2016)
Fogarty, C.B., Mikkelsen, M.E., Gaieski, D.F., Small, D.S.: Discrete optimization for interpretable study populations and randomization inference in an observational study of severe sepsis mortality. J. Am. Stat. Assoc. 111, 447–458 (2016)
Fogarty, C.B., Shi, P., Mikkelsen, M.E., Small, D.S.: Randomization inference and sensitivity analysis for composite null hypotheses with binary outcomes in matched observational studies. J. Am. Stat. Assoc. 112, 321–331 (2017)
Foster, E.M., Bickman, L.: Old wine in new skins: the sensitivity of established findings to new methods. Eval. Rev. 33, 281–306 (2009)
Frangakis, C.E., Rubin, D.B.: Principal stratification in causal inference. Biometrics 58, 21–29 (2002)
Franklin, J.M., Eddings, W., Glynn, R.J., Schneeweiss, S.: Regularized regression versus the high-dimensional propensity score for confounding adjustment in secondary database analyses. Am. J. Epidemiol. 182, 651–657 (2015)
Gastwirth, J.L.: Methods for assessing the sensitivity of comparisons in Title VII cases to omitted variables. Jurimetrics J. 33, 19–34 (1992)
Gastwirth, J.L., Krieger, A.M., Rosenbaum, P.R.: Dual and simultaneous sensitivity analysis for matched pairs. Biometrika 85, 907–920 (1998)
Gastwirth, J.L., Krieger, A.M., Rosenbaum, P.R.: Asymptotic separability in sensitivity analysis. J. R. Stat. Soc. B 62, 545–555 (2000)
Greenland, S.: Basic methods of sensitivity analysis. Int. J. Epidemiol. 25, 1107–1116 (1996)
Greevy, R., Lu, B., Silber, J.H., Rosenbaum, P.R.: Optimal matching before randomization. Biostatistics 5, 263–275 (2004)
Gross, D.B., Souleles, N.S.: Do liquidity constraints and interest rates matter for consumer behavior? Evidence from credit card data. Q. J. Econ. 117, 149–185 (2002)
Hahn, J.Y.: On the role of the propensity score in efficient semiparametric estimation of average treatment effects. Econometrica 66, 315–331 (1998)
Hamermesh, D.S.: The craft of labormetrics. Ind. Labor Relat. Rev. 53, 363–380 (2000)
Hansen, B.B.: The prognostic analogue of the propensity score. Biometrika 95, 481–488 (2008)
Haviland, A., Nagin, D.S., Rosenbaum, P.R.: Combining propensity score matching and group-based trajectory analysis in an observational study. Psychol. Methods 12, 247–267 (2007)
Haviland, A.M., Nagin, D.S., Rosenbaum, P.R., Tremblay, R.E.: Combining group-based trajectory modeling and propensity score matching for causal inferences in nonexperimental longitudinal data. Dev. Psychol. 44, 422–436 (2008)
Hirano, K., Imbens, G.W., Ridder, G.: Efficient estimation of average treatment effects using the estimated propensity score. Econometrica 71, 1161–1189 (2003)
Hill, A.B.: The environment and disease: association or causation? Proc. R. Soc. Med. 58, 295–300 (1965)
Hill, J.L. , Waldfogel, J. , Brooks-Gunn , J., Han, W.J.: Maternal employment and child development: a fresh look using newer methods. Dev. Psychol. 41, 833–850 (2005)
Ho, D.E., Imai, K., King, G., Stuart, E.A.: Matching as nonparametric preprocessing for reducing model dependence in parametric causal inference. Polit. Anal. 15, 199–236 (2007)
International Agency for Research on Cancer: IARC Monographs on the Valuation of Carcinogenic Risks of Chemicals to Humans: Chromium, Nickel and Welding, vol. 49, pp. 447–525. IARC, Lyon (1990)
Imai, K.: Statistical analysis of randomized experiments with non-ignorable missing binary outcomes: an application to a voting experiment. Appl. Stat. 58, 83–104 (2009)
Imbens, G.W.: The role of the propensity score in estimating dose response functions. Biometrika 87, 706–710 (2000)
Imbens, G.W.: Sensitivity to exogeneity assumptions in program evaluation. Am. Econ. Rev. 93, 126–132 (2003)
Imbens, G.W. : Nonparametric estimation of average treatment effects under exogeneity: a review. Rev. Econ. Stat. 86, 4–29 (2004)
Imbens, G.W., Wooldridge, J.M.: Recent developments in the econometrics of program evaluation. J. Econ. Lit. 47, 5–86 (2009)
Joffe, M.M., Ten Have, T.R., Feldman, H.I., Kimmel, S.E.: Model selection, confounder control, and marginal structural models: review and new applications. Am Stat. 58, 272–279 (2004)
Johnson, B.A., Tsiatis, A.A.: Estimating mean response as a function of treatment duration in an observational study, where duration may be informatively censored. Biometrics 60, 315–323 (2004)
Katan, M.B.: Commentary: Mendelian randomization, 18 years on. Int. J. Epidemiol. 33, 10–11 (2004)
Keele, L.J.: Rbounds: an R package for sensitivity analysis with matched data. http://www.polisci.ohio-state.edu/faculty/lkeele/rbounds.html
Lee, M.J., Lee, S.J.: Sensitivity analysis of job-training effects on reemployment for Korean women. Empir. Econ. 36, 81–107 (2009)
Lin, D.Y., Psaty, B.M., Kronmal, R.A.: Assessing sensitivity of regression to unmeasured confounders in observational studies. Biometrics 54, 948–963 (1998)
Lu, B., Rosenbaum, P.R.: Optimal matching with two control groups. J. Comput. Graph Stat. 13, 422–434 (2004)
Manski, C.: Nonparametric bounds on treatment effects. Am. Econ. Rev. 80, 319–323 (1990)
Manski, C.F. : Identification Problems in the Social Sciences. Harvard University Press, Cambridge (1995)
Manski, C.F., Nagin, D.S.: Bounding disagreements about treatment effects: a case study of sentencing and recidivism. Sociol. Method 28, 99–137 (1998)
Marcus, S.M.: Using omitted variable bias to assess uncertainty in the estimation of an AIDS education treatment effect. J. Educ. Behav. Stat. 22, 193–201 (1997)
McCaffrey, D.F., Ridgeway, G., Morral, A.R.: Propensity score estimation with boosted regression for evaluating causal effects in observational studies. Psychol. Methods 9, 403–425 (2004)
McKillip, J.: Research without control groups: a control construct design. In: Bryant, F.B., et al. (eds.) Methodological Issues in Applied Social Psychology, pp. 159–175. Plenum Press, New York (1992)
Meyer, B.D.: Natural and quasi-experiments in economics. J. Bus. Econ. Stat. 13, 151–161 (1995)
Mitra, N. Heitjan, D.F.: Sensitivity of the hazard ratio to nonignorable treatment assignment in an observational study. Stat. Med. 26, 1398–1414 (2007)
Normand, S-L., Landrum, M.B., Guadagnoli, E., Ayanian, J.Z., Ryan, T.J., Cleary, P.D., McNeil, B.J.: Validating recommendations for coronary angiography following acute myocardial infarction in the elderly: a matched analysis using propensity scores. J. Clin. Epidemiol. 54, 387–398 (2001)
Normand, S-L., Sykora, K., Li, P., Mamdani, M., Rochon, P.A., Anderson, G.M.: Readers guide to critical appraisal of cohort studies: 3. Analytical strategies to reduce confounding. Br. Med. J. 330, 1021–1023 (2005)
Pagano, M., Tritchler, D.: On obtaining permutation distributions in polynomial time. J. Am. Stat. Assoc. 78, 435–440 (1983)
Peel, M.J., Makepeace, G.H.: Differential audit quality, propensity score matching and Rosenbaum bounds for confounding variables. J. Bus. Financ. Account 39, 606–648 (2012)
Pimentel, S.D.: Large, sparse optimal matching with R package rcbalance. Obs. Stud. 2, 4–23 (2016)
Reynolds, K.D., West, S.G.: A multiplist strategy for strengthening nonequivalent control group designs. Eval. Rev. 11, 691–714 (1987)
Richardson, A., Hudgens, M.G., Gilbert, P.B., Fine, J.P.: Nonparametric bounds and sensitivity analysis of treatment effects. Stat. Sci. 29, 596–618 (2014)
Robins, J.M., Ritov, Y.: Toward a curse of dimensionality appropriate (CODA) asymptotic theory for semi-parametric models. Stat. Med. 16, 285–319 (1997)
Robins, J.M., Mark, S.D., Newey, W.K.: Estimating exposure effects by modeling the expectation of exposure conditional on confounders. Biometrics 48, 479–495 (1992)
Robins, J.M., Rotnitzky, A., Scharfstein, D.: Sensitivity analysis for selection bias and unmeasured confounding in missing data and causal inference models. In: Halloran, E., Berry, D. (eds.) Statistical Models in Epidemiology, pp. 1–94. Springer, New York (1999)
Rosenbaum, P.R.: Conditional permutation tests and the propensity score in observational studies. J. Am. Stat. Assoc. 79, 565–574 (1984)
Rosenbaum, P.R.: Dropping out of high school in the United States: an observational study. J. Educ. Stat. 11, 207–224 (1986)
Rosenbaum, P.R.: Sensitivity analysis for certain permutation inferences in matched observational studies. Biometrika 74, 13–26 (1987)
Rosenbaum, P.R.: The role of a second control group in an observational study (with Discussion). Stat. Sci. 2, 292–316 (1987)
Rosenbaum, P.R.: Model-based direct adjustment. J. Am. Stat. Assoc. 82, 387–394 (1987)
Rosenbaum, P.R.: The role of known effects in observational studies. Biometrics 45, 557–569 (1989)
Rosenbaum, P.R.: On permutation tests for hidden biases in observational studies. Ann. Stat. 17, 643–653 (1989)
Rosenbaum, P.R.: Hodges-Lehmann point estimates in observational studies. J. Am. Stat. Assoc. 88, 1250–1253 (1993)
Rosenbaum, P.R.: Quantiles in nonrandom samples and observational studies. J. Am. Stat. Assoc. 90, 1424–1431 (1995)
Rosenbaum, P.R.: Signed rank statistics for coherent predictions. Biometrics 53, 556–566 (1997)
Rosenbaum, P.R.: Choice as an alternative to control in observational studies (with Discussion). Stat. Sci. 14, 259–304 (1999)
Rosenbaum, P.R.: Stability in the absence of treatment. J. Am. Stat. Assoc. 96, 210–219 (2001)
Rosenbaum, P.R.: Observational Studies, 2nd edn. Springer, New York (2002)
Rosenbaum, P.R.: Covariance adjustment in randomized experiments and observational studies (with Discussion). Stat. Sci. 17, 286–327 (2002)
Rosenbaum, P.R.: Design sensitivity in observational studies. Biometrika 91, 153–164 (2004)
Rosenbaum, P.R.: Heterogeneity and causality: unit heterogeneity and design sensitivity in observational studies. Am Stat. 59, 147–152 (2005)
Rosenbaum, P.R.: Sensitivity analysis for m-estimates, tests, and confidence intervals in matched observational studies. Biometrics 63, 456–464 (2007)
Rosenbaum, P.R.: Design sensitivity and efficiency in observational studies. J. Am. Stat. Assoc. 105, 692–702 (2010)
Rosenbaum, P.R.: Optimal matching of an optimally chosen subset in observational studies. J. Comput. Graph Stat. 21, 57–71 (2012)
Rosenbaum, P.R.: Observation and Experiment: An Introduction to Causal Inference. Harvard University Press, Cambridge, MA (2017)
Rosenbaum, P.R.: Sensitivity analysis for stratified comparisons in an observational study of the effect of smoking on homocysteine levels. Ann. Appl. Stat. 12, 2312–2334 (2018)
Rosenbaum, P.R.: Modern algorithms for matching in observational studies. Ann. Rev. Stat. Appl. 7, 143–176 (2020)
Rosenbaum, P.R., Krieger, A.M.: Sensitivity analysis for two-sample permutation inferences in observational studies. J. Am. Stat. Assoc. 85, 493–498 (1990)
Rosenbaum, P.R., Rubin, D.B.: The central role of the propensity score in observational studies for causal effects. Biometrika 70, 41–55 (1983)
Rosenbaum, P.R., Rubin, D.B.: Assessing sensitivity to an unobserved binary covariate in an observational study with binary outcome. J. R. Stat. Soc. B45, 212–218 (1983)
Rosenbaum, P.R., Rubin, D.B.: Reducing bias in observational studies using subclassification on the propensity score. J. Am. Stat. Assoc. 79, 516–524 (1984)
Rosenbaum, P.R., Rubin, D.B. : Constructing a control group by multivariate matched sampling methods that incorporate the propensity score. Am. Stat. 39, 33–38 (1985)
Rosenbaum, P.R., Rubin, D.B.: The bias due to incomplete matching. Biometrics 41, 106–116 (1985)
Rosenbaum, P.R., Silber, J.H.: Sensitivity analysis for equivalence and difference in an observational study of neonatal intensive care units. J. Am. Stat. Assoc. 104, 501–511 (2009)
Rosenbaum, P.R., Silber, J.H.: Amplification of sensitivity analysis in observational studies. J. Am. Stat. Assoc. 104, 1398–1405 (2009)
Rosenbaum, P.R., Small, D.S.: An adaptive Mantel-Haenszel test for sensitivity analysis in observational studies. Biometrics 73, 422–430 (2017)
Rosenzweig, M.R., Wolpin, K.I.: Natural ‘natural experiments’ in economics. J. Econ. Lit. 38, 827–874 (2000)
Rotnitzky, A., Robins, J.M.: Semiparametric regression estimation in the presence of dependent censoring. Biometrika 82, 805–820 (1995)
Rubin, D.B., Thomas, N.: Characterizing the effect of matching using linear propensity score methods with normal distribution. Biometrika 79, 797–809 (1992)
Rubin, D.B., Thomas, N.: Combining propensity score matching with additional adjustments for prognostic covariates. J. Am. Stat. Assoc. 95, 573–585 (2000)
Rudolph, K.E., Stuart, E.A.: Using sensitivity analyses for unobserved confounding to address covariate measurement error in propensity score methods. Am. J. Epidemiol. 187, 604–613 (2017)
Rutter, M.: Identifying the Environmental Causes of Disease: How do We Decide What to Believe and When to Take Action? Academy of Medical Sciences, London (2007)
Shadish, W.R., Cook, T.D.: The renaissance of field experimentation in evaluating interventions. Annu. Rev. Psychol. 60, 607–629 (2009)
Shadish, W.R., Cook, T.D., Campbell, D.T.: Experimental and Quasi-Experimental Designs for Generalized Causal Inference. Houghton-Mifflin, Boston (2002)
Shepherd, B.E., Gilbert, P.B., Jemiai, Y., Rotnitzky, A.: Sensitivity analyses comparing outcomes only existing in a subset selected post-randomization, conditional on covariates, with application to HIV vaccine trials. Biometrics 62, 332–342 (2006)
Shepherd, B.E., Gilbert, P.B., Mehrotra, D.V.: Eliciting a counterfactual sensitivity parameter. Am Stat. 61, 56–63 (2007)
Silber, J.H. , Rosenbaum, P. R., Trudeau, M.E. , Chen, W. , Zhang, X. , Lorch, S.L. , Rapaport-Kelz, R. , Mosher, R.E. , Even-Shoshan, O. : Preoperative antibiotics and mortality in the elderly. Ann. Surg. 242, 107–114 (2005)
Small, D., Rosenbaum, P.R.: War and wages: The strength of instrumental variables and their sensitivity to unobserved biases. J. Am. Stat. Assoc. 103, 924–933 (2008)
Spielman, R.S., Ewens, W.J.: A sibship test for linkage in the presence of association: the sib transmission/disequilibrium test. Am. J. Hum. Genet. 62, 450–458 (1998)
Stolley, P.D.: When genius errs—R.A. Fisher and the lung cancer controversy. Am. J. Epidemiol. 133, 416–425 (1991)
Stone, R.: The assumptions on which causal inferences rest. J. R. Stat. Soc. B 55, 455–466 (1993)
Traskin, M., Small, D.S.: Defining the study population for an observational study to ensure sufficient overlap: a tree approach. Stat. Biosci. 3, 94–118 (2011)
Trochim, W.M.K.: Pattern matching, validity and conceptualization in program evaluation. Eval. Rev. 9, 575–604 (1985)
Vandenbroucke, J.P.: When are observational studies as credible as randomized trials? Lancet 363, 1728–1731 (2004)
VanderWeele, T.: The use of propensity score methods in psychiatric research. Int. J. Methods Psychol. Res. 15, 95–103 (2006)
Volpp, K.G., Rosen, A.K., Rosenbaum, P.R., Romano, P.S., Even-Shoshan, O., Wang, Y., Bellini, L., Behringer, T., Silber, J.H.: Mortality among hospitalized Medicare beneficiaries in the first 2 years following ACGME resident duty hour reform. J. Am. Med. Assoc. 298, 975–983 (2007)
Wang, L.S., Krieger, A.M.: Causal conclusions are most sensitive to unobserved binary covariates. Stat. Med. 25, 2257–2271 (2006)
Weiss, N.S.: Can the “specificity” of an association be rehabilitated as a basis for supporting a causal hypothesis? Epidemiology 13, 6–8 (2002)
Werfel, U., Langen, V., Eickhoff, I., Schoonbrood, J., Vahrenholz, C., Brauksiepe, A., Popp, W., Norpoth, K.: Elevated DNA single-strand breakage frequencies in lymphocytes of welders. Carcinogenesis 19, 413–418 (1998)
Wintemute, G.J. , Wright, M.A., Drake, C.M. , Beaumont, J.J.: Subsequent criminal activity among violent misdemeanants who seek to purchase handguns: risk factors and effectiveness of denying handgun purchase. J. Am. Med. Assoc. 285, 1019–1026 (2001)
Wolfe, D.A.: A characterization of population weighted symmetry and related results. J. Am. Stat. Assoc. 69, 819–822 (1974)
Wyss, R., Schneeweiss, S., van der Laan, M., Lendle, S.D., Ju, C., Franklin, J.M.: Using super learner prediction modeling to improve high-dimensional propensity score estimation. Epidemiology 29, 96–106 (2018)
Yu, B.B., Gastwirth, J.L.: Sensitivity analysis for trend tests: application to the risk of radiation exposure. Biostatistics 6, 201–209 (2005)
Zanutto, E., Lu, B., Hornik, R.: Using propensity score subclassification for multiple treatment doses to evaluate a national antidrug media campaign. J. Educ. Behav. Stat. 30, 59–73 (2005)
Zhao, Q.: On sensitivity value of pair-matched observational studies. J. Am. Stat. Assoc. 114, 713–722 (2019)
Zhao, Q.: Covariate balancing propensity score by tailored loss functions. Ann. Stat. 47, 965–993 (2019)
Zhao, Q., Small, D.S., Bhattacharya, B.B.: Sensitivity analysis for inverse probability weighting estimators via the percentile bootstrap. J. R. Stat. Soc. B 81, 735–761 (2019)
Zubizarreta, J.R., Cerda, M., Rosenbaum, P.R.: Effect of the 2010 Chilean earthquake on posttraumatic stress: reducing sensitivity to unmeasured bias through study design. Epidemiology 7, 79–87 (2013)
Author information
Authors and Affiliations
Appendix: Exact Computations for Sensitivity Analysis
Appendix: Exact Computations for Sensitivity Analysis
For small to moderate I, the exact upper bound on the distribution of Wilcoxon’s signed rank statistic may be obtained quickly in R as the convolution of I probability generating functions. Only the situation without ties is considered. Pagano and Tritchler [73] observed that permutation distributions may often be obtained in polynomial time by applying the fast Fourier transform to the convolution of characteristic functions or generating functions. Here, probability generating functions are used along with the R function convolve.
The distribution of \(\overline {\overline {T}}\) is the distribution of the sum of I independent random variables, i = 1, 2, …, I, taking the value i with probability \(\varGamma /\left ( 1+\varGamma \right ) \) and the value 0 with probability \(1/\left ( 1+\varGamma \right ) \). The ith random variable has probability generating function
and \(\overline {\overline {T}}\) has generating function \(\varPi _{i=1}^{I}h_{i}\left ( x\right ) \). In R, the generating function of a random variable taking integer values 0, 1, …, B is represented by a vector of dimension B + 1 whose b + 1 coordinate gives the probability that the random variable equals b. For instance, \(h_{3}\left ( x\right ) \) is represented by
The distribution of \(\overline {\overline {T}}\) is obtained by convolution as a vector with \(1+I\left ( I+1\right ) /2\) coordinates representing \(\varPi _{i=1}^{I}h_{i}\left ( x\right ) \) and giving \(\Pr \left ( \overline { \overline {T}}=b\right ) \) for b = 0, 1, …, \(I\left ( I+1\right ) /2\), where \(I\left ( I+1\right ) /2=\sum i\).
Rights and permissions
Copyright information
© 2020 Springer Nature Switzerland AG
About this chapter
Cite this chapter
R. Rosenbaum, P. (2020). Two Simple Models for Observational Studies. In: Design of Observational Studies. Springer Series in Statistics. Springer, Cham. https://doi.org/10.1007/978-3-030-46405-9_3
Download citation
DOI: https://doi.org/10.1007/978-3-030-46405-9_3
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-030-46404-2
Online ISBN: 978-3-030-46405-9
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)